HYPOTHETICAL PERFORMANCE These are paper-trading portfolios. No client money is invested, no orders are routed to any venue, and no result shown here was achieved with real capital. Hypothetical results have inherent limitations and do not reflect the effect of real order flow on price.
Transparent AI Investment Platform
Home / methodology

How this works

MAR Article 4 requires a summary of the basis of valuation, the methodology and the underlying assumptions, an indication of any change to them, and a sensitivity analysis. This page is that, and it says where it falls short.

The single most important architectural decision

The AI ranks and explains. Deterministic code decides everything that touches money.

The model's output schema has no field for a price, a share count, a weight, a notional or a risk limit. It expresses direction and conviction, plus a relative stop preference — "tight", "standard" or "wide" — which the mandate clamps. Everything quantitative is computed downstream from the mandate and from real prices.

This is not a prompt instruction, which a model can ignore. It is the type system, and a test fails the build if anyone widens the schema.

Three gates the AI cannot see or bypass

Gate 1 — proposal validation

Is this instrument in the universe? Is the action permitted by this mandate? Is the portfolio allowed to open new positions right now? Every refusal carries a stable reason code and is published.

Gate 2 — sizing

The share quantity is computed from the mandate, never from the model. Position caps, sector caps, issuer caps, sleeve bands and order-size-versus-volume limits all bind here.

Gate 3 — portfolio state

Loss limits, drawdown ladders, volatility ladders, stale-data halts and circuit breakers. A halted portfolio may still reduce risk — exits always run, because a cap that stopped a stop firing would turn a bookkeeping limit into an unbounded loss.

Fills, and why look-ahead is impossible rather than avoided

A fill is struck on a volume-weighted average price over a window that begins strictly after the decision timestamp plus a latency delay. No fill can use information from the bar in which the decision was made.

Conventions that would allow it are banned outright rather than discouraged — the paper broker raises on them and a test asserts that no mandate uses one:

Banned conventionWhy
same_bar_closeThe decision used data up to the close and would fill at the close. Indefensible.
same_bar_vwapOverlaps the decision window.
bar_high_or_lowUses information unavailable at decision time.
next_day_open_for_intraday_decisionInjects a full overnight gap unrelated to the strategy.

Costs

Every result is net of modelled execution costs, itemised rather than collapsed into one opaque number: spread, market impact, commission, SEC Section 31 fees and FINRA trading activity fees. Gross and net are always reported side by side so the cost drag is visible.

Published verbatim, including the assumptions that are not sourced. Where a number is a guess it is labelled a guess. A cost model that hides its weak assumptions makes every result downstream of it unfalsifiable.

Broker profiles modelled

zero_commission_us · swiss_retail · institutional · crypto_retail_punitive · intraday_punitive

Section 31 fee rates are taken from an effective-date table rather than hardcoded — the rate was $0.00 for part of fiscal 2026 and then $20.60 per million, and a single hardcoded value would be wrong on both sides of the change.

Sensitivity analysis — partially built

Status: partial. Execution-cost sensitivity across 5 broker profiles (zero_commission_us, swiss_retail, institutional, crypto_retail_punitive, intraday_punitive) is computed and published.

Rendering every performance figure under all profiles simultaneously, and a second axis at +/-50% on the spread assumptions, is specified but not yet built. This gap is published rather than omitted.

Statistics that are deliberately withheld

Annualised return, volatility, Sharpe, Sortino and Calmar are not computed below sixty observations, and the site shows "withheld" rather than a number. Annualising a three-week return produces a figure that looks precise and is not, and publishing one would be a small lie repeated on every page.

A benchmark comparison is aligned on overlapping dates only. A benchmark series that started later than the portfolio would flatter it.

Changes to any of this

Rulebook changes are events in the same hash chain as trades, carrying a full diff, a reason and an effective date. Model and prompt changes are the same. You cannot quietly loosen a limit after a bad quarter here, because the loosening would appear in sequence next to the quarter that prompted it. Current rulebook versions:

P01
v1.1.1
P02
v2.0.1
P03
v1.1.1
P04
v1.1.1
P05
v1.0.0
P06
v1.1.1
P07
v1.2.1
P08
v1.2.1
P09
v1.1.0

What the actions mean

MAR Article 4 asks that the meaning of a recommendation be explained. The permitted actions here are not price targets or ratings. BUY and ADD express that an instrument should be entered or increased, subject to the mandate computing whether and by how much. TRIM and SELL express reduction or exit. HOLD is an explicit no-change. FLATTEN closes everything. TIGHTEN_STOP moves a stop closer, within limits the mandate sets. FLAG_THESIS_BREAK and FLAG_DATA_ANOMALY raise something for attention without proposing a trade at all.

Follow the record.Every decision, fill and refusal, as it is published — including the ones that lose money.Join on Telegram →

Crypto market data Powered by CoinGecko · Logos provided by Logo.dev · Company logos identify holdings only and imply no endorsement.